AI Boosts Lesion Detection in IBD, Outperforming Conventional Methods
BERLIN, Germany — A groundbreaking, multicenter study has shown that artificial intelligence (AI)-assisted capsule endoscopy (CE) readings demonstrate superior sensitivity and accuracy in identifying ulcers and erosions in patients with inflammatory bowel disease (IBD) compared to standard reading methods. Beyond increasing diagnostic accuracy, the AI model dramatically reduced exam reading times.
Additionally, the study validated an AI model for small-bowel CE in real clinical scenarios. The AI model tackles long-standing limitations of CE, including time-intensive readings and variability between observers.
“It’s a huge improvement on the technology readiness level of the AI model,” stated Miguel Mascarenhas, MD, PhD, Head of the Precision Medicine Unit at Hospital São João, Faculty of Medicine of the University of Porto, Porto, Portugal, and senior study investigator. He explained this is the first AI system using a CE platform that has proven so effective across numerous real-world clinical settings. “This technology is set to transform endoscopic practice and clinical management in inflammatory bowel disease.”
The findings (abstract DOP089) were presented at the European Crohn’s and Colitis Organisation 2025 Congress by Francisco Mendes, MD, a resident in gastroenterology at Hospital São João.
More Lesions Detected, in Less Time
The prospective study was conducted between January 2021 and April 2024, involving several centers in Portugal, Spain, and the United States. Researchers evaluated the performance of two CE devices (PillCamSB3 and Olympus EC-10) using 137 CE exams from 137 patients, 49 of whom had Crohn’s disease. The study compared AI-assisted readings with those obtained through the standard of care, with expert board consensus serving as the gold standard.
Key performance metrics included sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). During expert board review, ulcers and erosions were detected in 56 patients (40.9%), with a sensitivity of 60.7%, a specificity of 98.8%, a PPV of 97.1%, and an NPV of 78.4%, resulting in an overall accuracy of 83.2% for the detection of ulcers and erosions.
In comparison, the AI-assisted readings exceeded conventional readings with a sensitivity of 94.6%, a specificity of 80.2%, a PPV of 76.8%, and an NPV of 95.6%, resulting in an overall accuracy of 86.1%.
The AI-assisted model’s diagnosis was found to be noninferior (P < .001) and superior (P < .001) to conventional diagnosis for detecting ulcers and erosions. The AI model showed consistent performance across various CE devices and participating centers. Furthermore, the average reading time per AI-assisted exam was under 4 minutes (239 seconds), compared to around 1.0-1.5 hours using standard-of-care readings.
Mascarenhas noted that the improved diagnostic accuracy of this AI model, achieved in significantly less time, enables clinicians to dedicate more time to patient interaction and other essential care-related tasks. He added that CE holds considerable promise not only in IBD but also in other GI-related screenings, including colorectal cancer screening. He believes that once the reading time bottleneck is resolved, CE will become the first-line tool for screening.
Shomron Ben-Horin, MD, Director, Sheba Medical Center, Tel-Aviv University, Tel-Aviv, Israel, told Medscape Medical News that the reading time is “one of several barriers” to integrating CE into clinical practice. However, he expressed that it “is the most accurate modality for detection of inflammatory activity along the entire small bowel.”
Based on the study results, AI is the best approach to the problem, according to Ben-Horin, who wasn’t involved in the study. “There was even a signal for better accuracy, which is intriguing,” he added. This study suggests that AI is better than physicians at reading, which is important.
Also commenting was Miles Parkes, MD, consultant gastroenterologist at Addenbrooke’s Hospital in Cambridge, United Kingdom. He told Medscape Medical News, that both the sensitivity and the specificity of the output are helpful, but there could be some details that need further examination. “However, as a general principle, the performance of this model is impressive.”
Mascarenhas and Mendes reported no financial disclosures. Ben Horin received fees from Medtronic to attend the conference. Parkes reported no financial disclosures.
